Greenfield Access Network
khelms at zcorum.com
Thu Jul 31 17:59:52 UTC 2014
On Thu, Jul 31, 2014 at 12:07 PM, Colton Conor <colton.conor at gmail.com>
> Thanks for the long post.
> We will use a layer 2 10G aggregation switch then to aggregate the chassis
> at the core location. Do you have any recommendations on 10G switches?
Not really, just stick with one of the major brands and you _should_ be
> Yes I realize the math is a little backwards as this is all hypothetical
> at this point. We would provision each ONT as a shared 1Gbps offering
> similar to Google Fiber. We know there will be a large amount of
> oversubscription as no one really uses a full Gbps or anywhere close to it.
> I just wanted to stress the point that carrier redundancy at the 10G level
> would be a requirement for the core router, and it should of course have
> 10G links going to the uplinks on the aggregation switch. I think the Cisco
> ASR9k and the Juniper MX line will do well. Not sure if there are any
> others that can handle this level of traffic on the BGP side?
That's reasonable IMO and yes, I think the Juniper MX can handle that as
well as some other functions for you related to subscriber management if
you want. The MX line has a full BRAS set of capabilities built into it
that it inherited from the older ERX line, but they're commonly deployed
without using any of them of as well.
> So we have a 10G aggregation switch to aggregate the chassis uplink
> connections, and a 10G router BGP capable router.
> I really liked your article on DHCP vs PPP for DSL networks. We definitely
> agree the way to go is with a DHCP server. A couple of items your article
> left as big questions:
> 1. The article mentioned DHCP doesn't do the other part of what PPPoE or
> PPPoA does, which is generate RADIUS accounting records that give us the
> bandwidth information. So that’s one of the main challenges in switching to
> a DHCP based system. So, how do you handle bandwidth tracking in an all
> DHCP environment then? If I want to track how many GB a customer used last
> month, or the average Mbps used how do you do so?
There are a few ways to get at that problem. You can use Netflow/IPFIX
collection to gather the usage from your router, accepting that you're only
going to get information on layer 3 traffic, which generally isn't a
problem. You will need to match the IPs up against your Option 82 parsing
which will give you the circuit ID, IP address, and WAN MAC of the ONT.
You can also poll your shelves via SNMP, CLI, TL-1, and/or Netconf to
collect the data and put it into a database in much the same way you can
use RADIUS accounting data.
> 2. I liked your option 82 example, and that works well for DSL networks
> where one port is tied to one customer. But how does option 82 work when
> you have multiple customers hanging off a GPON port? What does GPON use a
> subport identifier?
Yep, the different vendors implement it slightly differently, usually the
ONT MAC/serial will be included or the ONT ID will be included. Talk with
your vendor, all the major OLT vendors are very familiar with Option 82 and
in many cases they can tailor what their boxes send to make it easier for
> 3. You mentioned, DHCP is again, not a authentication protocol. So what
> handles authentication then if only DHCP is used, and there are no
> usernames and passwords? I guess for DSL networks you can enable or disable
> the port to allow or disallow access, and Option 82 for identification? I
> assume you wouldn't want to shut off the GPON OLT port if one customer
> wasn't paying their bill as it would affect the other customers on that
> port. I assume access vendors allow you to shut down the sub port or ONT in
> this situation for GPON? Still that seems messy having to login to a shelf
> or EMS system or API to an EMS system especially if you have multiple
> access vendors in a network. Is there a way to do authentication with DHCP?
> What about open networks like wifi where anyone can connect, so you don't
> have the ability to turn of the port or disable the end device?
> 4. I don't think anyone is buying a BRAS anymore, but looks like Cisco,
> Juniper, and ALU have what they call BGN, Broadband Subscriber Management,
> and other similar software. How are these different from BRAS functionality?
First, if you can manage it turn on DOCSIS provisioning of your GPON
network. AFAIK only Calix has announced this functionality, but I expect
the others to follow suit now that there is an official effort at CableLabs
to allow that.
The notion of managing ports and profiles via (an ever changing) shelf API
is one of the main reasons that telco billing systems cost so much compared
to cable billing systems. If you can't swing DPoG then you're kind of
stuck, either you can implement the API your vendor supplies with your
billing system, manage the profile assignment manually (yuck), or just
provision everyone with the same speed (only works for data only
deployments), or go down the older route of putting in a BRAS and making
sure the ONTs you're deploying have a PPPoE client embedded in them.
Having to deploy an external router for each customer, which I've seen
some operators do, makes your install costs higher and makes
> So it looks like there are open source and commercial solutions for DHCP
> and DNS. Some providers like Infloblox seems to integrate all these into
> So if we have a core router that speaks BGP, a 10G aggregation switch to
> aggregate the the chassis, and a device like Infloblox or the other
> commercial solutions you mentioned that do DHCP/DNS, is there anything else
> that is needed besides the access gear already mentioned in the
> assumptions? Are these large and expensive commercial BGN/Broadband
> Subscriber management products a thing of the past or still very relevant
> in todays environment?
They're not very relevant, once the the OLT vendors realized they could
snoop the DHCP session and enforce what the server provided the need for
subscriber management pieces really dropped. You've listed the bare
essentials for a functional network, there are lots of things that are
helpful or useful, but what you have is functional. Having said all of
that, there are some relatively unobtrusive ways to have some level of
authentication, I just don't think they're very valuable.
> On Thu, Jul 31, 2014 at 8:54 AM, Scott Helms <khelms at zcorum.com> wrote:
>> "What is the ideal way to aggregate the 40 10G connections from the
>> of the chassis? I would guess a 10G switch since 10G ports on a router
>> would be much more expensive?"
>> Definitely aggregate into a switch first unless you want to run a Layer 3
>> switch as your router, which I don't recommend.
>> "Which router is recommended to handle 4 10G internet connections with
>> tables, and then at least 4 10G ports going back to the 10G aggregation
>> Your math is a little backwards, its very unlikely that you're going to
>> have 40 Gbps of Internet (or other interconnection) for the router to
>> actually have to process. What is the average provisioned speed for each
>> of the 10k PON ports? What over subscription rate are you planning for?
>> What, if anything, will you be carrying on net, ie bandwidth consumption
>> that won't come from or go to the public Internet? Your own video, voice,
>> or other service are examples of things that are often on net. In any case
>> you're probably in the ASR family with Cisco and I can't remember
>> the equivalent from Juniper.
>> How do you handle IP address management? a /20 is only 4096 IP addresses,
>> but the network would have potentially 10,000 customers. Assume that
>> getting more space from ARIN is not an option. Is CGN an option?
>> CGN is the option of last resort IMO, but you may have to consider it. A
>> better approach is to see if your backbone providers will agree to give
>> some blocks that you can announce and use those blocks for dynamic
>> customers only. Your static IP customers should come from your direct ARIN
>> allotment in case you need to choose a new backbone provider, which is
>> extremely common over time.
>> "Dynamic IP
>> addresses? DHCP?"
>> DHCP with enforcement from the shelves. All the major OLT vendors
>> support doing this so that a customer can only use the address assigned to
>> him by DHCP and nothing else, except for those customers that you choose to
>> hard code. Make most of your "static" customers actually DHCP reservations
>> and only hard code those that you must.
>> "How do you separate users and traffic? VLANs, Service VLANs, Per Customer
>> VLANs, Usernames? Passwords? PPPoE? MAC Separation?
>> Is a BRAS or BGN functionally really needed or are these older concepts?"
>> DHCP, with Option 82 logging for the circuit ID is the better path than a
>> BRAS (PPPoE) these days. Here's a paper we put together on that topic a
>> while back:
>> Depending on your OLT vendor you can either use their built in port
>> isolation or QinQ tagging, both are reliable and scalable, just ask your
>> vendor which is the best option for your specific gear.
>> "If CGNAT or DHCP is needed, what will host the CGNAT or DHCP service? The
>> core router, a linux box, or something else?"
>> I wouldn't have those two services connected personally, though there are
>> hooks for some of the CGN boxes to talk to DHCP servers. I would hope you
>> can get another 6k addresses and avoid the need for CGN altogether. Having
>> said that, have you tested your OLTs and ONTs for IPv6 interoperability?
>> If they don't handle it well then you're going to have to think about
>> alternatives like 6RD (http://en.wikipedia.org/wiki/IPv6_rapid_deployment
>> For DHCP at your scale you can run ISC DHCP (
>> http://www.isc.org/downloads/dhcp/) which is the most common open source
>> DHCP daemon if you someone who can take care of a Linux server, parse the
>> Option 82 information for logging, and handle the configuration of the DHCP
>> daemon itself. Otherwise you might want to look at commercial products
>> designed for the service provider market like Incongito's BCC and Cisco's
>> BAC (CNR replacement)
>> "What about DNS?
>> Is a firewall needed in the core?
>> What else is needed?"
>> There are two kinds of DNS, caching (recursive) and authoritative. The
>> first is what your customers will use to resolve things on the Internet and
>> the second is used to provide caching name servers on the Internet with
>> information about domains you control (are authoritative for). The first
>> needs good performance, availability, and scalability since your customers
>> will use your caching name servers constantly. Most people can run BIND at
>> your scale, again if you have someone with Linux experience, but there are
>> other alternatives. PowerDNS has both caching and authoritative modules
>> and there are some commercial offerings out there both as cloud hosting and
>> local deployments. Your backbone provider will also often have caching
>> name servers your customers can use, but the quality varies quite a bit.
>> You can also, especially at first, leverage some of the free offerings
>> like Google's DNS. I don't recommend firewalls for service provider
>> networks, but you should make sure your gear can run (and is configured to
>> do so) BCP 38.
>> Scott Helms
>> Vice President of Technology
>> (678) 507-5000
>> On Thu, Jul 31, 2014 at 9:23 AM, Colton Conor <colton.conor at gmail.com>
>>> If a new operator or city is building a greenfield access network from
>>> ground up, what software and hardware is needed in the core network to
>>> provide and manage residential and business internet services similar to
>>> the likes of AT&T, Comcast, and Google Fiber? Television and Telephone
>>> services are not to be considered only internet.
>>> Assume hypothetically the operator already has the following in place:
>>> 10 GPON OLTs Chassis from an access vendor in 10 POPs around town (each
>>> has 1 Chassis). Each OLT Chassis has 4 10G Uplinks back to the core.
>>> Dark fiber going from the POP locations back to the core location
>>> Assume a 32:1 way split, and each OLT chassis has enough ports populated
>>> serve the area.
>>> 10,000 GPON ONTs. The ONTs can be put in routed gateway or bridged mode.
>>> Assume you are building a network designed to serve 10,000 subs
>>> All the fiber splitters, ducts, fiber, etc connecting the OLTs to the
>>> is already in place
>>> ASN from ARIN
>>> /20 of IPv4 space and /32 of IPv6 space from ARIN
>>> 4 burstable 10G internet connections from 4 tier 1 internet providers
>>> Questions are:
>>> What is the ideal way to aggregate the 40 10G connections from the
>>> of the chassis? I would guess a 10G switch since 10G ports on a router
>>> would be much more expensive?
>>> Which router is recommended to handle 4 10G internet connections with
>>> tables, and then at least 4 10G ports going back to the 10G aggregation
>>> How do you handle IP address management? a /20 is only 4096 IP addresses,
>>> but the network would have potentially 10,000 customers. Assume that
>>> getting more space from ARIN is not an option. Is CGN an option? Dynamic
>>> addresses? DHCP?
>>> How do you separate users and traffic? VLANs, Service VLANs, Per Customer
>>> VLANs, Usernames? Passwords? PPPoE? MAC Separation?
>>> Is a BRAS or BGN functionally really needed or are these older concepts?
>>> If CGNAT or DHCP is needed, what will host the CGNAT or DHCP service? The
>>> core router, a linux box, or something else?
>>> What about DNS?
>>> Is a firewall needed in the core?
>>> What else is needed?
>>> Is there a guide out there somewhere? I know many cities are looking at
>>> building their own network, and have similar questions. Access vendors
>>> willing to sell gear all day long, but then they leave it up to the
>>> operator/city to answer these harder questions.
>>> How would you build a access network from the ground up if you had the
>>> resources and time to do so? Would you even use GPON? Even if GPON was
>>> used and another access technology like AE, VDSL2, or wireless was used I
>>> think many of these questions would be the same.
More information about the NANOG